Benchmark report
Task verified
13m 8s elapsed
The saved attempt met the documented success criterion.
100/ 100
Agent-Ready- SInstall
- SAuth
- SExecute
- SDocs
- SCurrent
Nothing substantive: the quickstart omitted raw HTTP endpoint details, costing one extra doc fetch, and the Task API run itself took ~25s to complete.
Where the time went
13:08 agent time- Starting1:39
- Research1:04
- Attempt6:50
- Verify3:35
Judge reason
Saved evidence proves a live, authenticated call to Parallel's documented answer operation. main.py uses the official parallel-web 1.3.7 SDK (client.task_run.create with task_spec output_schema type 'text', processor 'core', then client.task_run.result) against POST https://api.parallel.ai/v1/tasks/runs, with the key read from a private file via PARALLEL_API_KEY_FILE (x-api-key path). response.json holds the full server response: run_id trun_cee34ffc8fd64f6da3ec0afc411e8b0c, status 'completed', created_at 2026-11-10T20:55:27Z / modified_at 20:55:52Z, output.content stating "PostgreSQL's default transaction isolation level is **READ COMMITTED**", and output.basis[0].citations with two absolute HTTPS URLs (postgresql.org/docs/current/transaction-iso.html and .../sql-set-transaction.html) plus verbatim source excerpts. run-output.txt (tee'd from the actual run at transcript idx 39-42) prints the same answer text and both citation URLs, matching the saved response exactly; nothing was mocked or hard-coded, and the answer came from output.content rather than any local model. Beta check was done correctly: docs confirm POST /v1/tasks/runs is not labeled beta (the 'Chat API (Beta)' tag is a different API) and the only beta marking, the legacy parallel-beta: field-basis-2025-11-25 header, is documented as unnecessary and was not sent.
Attempt record
- 4:48Answered from saved key: Parallel API key (PARALLEL_API_KEY)no wait
- —First success verified
Attempt artifacts saved.